在人們越來越重視個人隱私保障的時代，各方無不努力的構思出新的身分認證方式，來確保使用者的身分可以被驗證而不會被盜用，而傳統上，有許多方法是透過每個人本身自帶的生物特徵的唯一性來進行身分辨識，諸如，掌紋、指紋、虹膜等等的單一生物特徵，以單一生物特徵進行身分辨識的準確性仍有成長空間；同時，近年深度學習網路隨著高性能計算機技術的進步，已運用於各領域之研究；因此本文研究應 用深度學習神經網路於多頻譜手掌影像的多模式生物辨識，以期提高身分辨識的準確性。本文以 CAISA 多光譜手掌影像資料集為實驗基礎，每種光譜的每一張手掌影像中包含掌紋、手型、指節紋路及指紋等多模式生物特徵，將手掌的多模式生物特徵輸入深度學習網路，運用手掌影像中每一種生物特徵，以便提高辨識的正確性。為了有效的增進訓練網路模型的精確度，首先嘗試過未經預訓練的網路模型進行訓練及測試，發現成果不如預期時；導入預訓練網路模型作為一種增進辨識精確度的作法，訓練後得到的精確度大幅上升，同時透過各類型不同資料擴增方法來進一步驗證各個預訓練網路模型對於不同變量下的適應能力，於實驗之中，使用了隨機旋轉、平移及亮度變化等常見的攝影環境改變進行資料擴增，從而發現此舉除了增強網路模型的精確度之外，也增強了網路模型適應影像變化的能力，實驗結果驗證了透過使用適當預訓練深度學習神經網路模型，以及合適的資料擴增方法增加資料量，以多光譜之手掌多模式特徵進行身分辨識具有很大的發展潛力。 ;In an age when people are more and more concerned about personal privacy protection, all parties are working hard to conceive new ways of identity authentication to ensure that the user′s identity can be verified and will not be stolen. Therefore, this paper investigates the application of deep learning neural networks to multi-modal biometric recognition of multi spectral palm images in order to improve the accuracy of body recognition. In this paper, we use the CASIA multi-spectral palm image dataset as the basis of the experiment. Each palm image of each spectrum contains multi-modal biometric features such as palm print, hand shape, knuckle pattern, and fingerprint, etc. The multi-modal biometric features of the palm are input into the deep learning network, and each biometric feature of the palm image is applied to improve the accuracy of recognition. In order to effectively improve the accuracy of the trained models, we first tried to train and test the unpre-trained network models and found that the results were not as good as expected; we introduced the pre-trained models as a way to improve the accuracy of recognition, and the accuracy obtained after training increased significantly. In the experiments, data augmentation was performed using common camera environment changes such as random rotation, panning, and luminance changes, and it was found that this not only enhanced the accuracy of the model, but also enhanced the ability of the model to adapt to image changes. The experimental results validate the potential of using multispectral palm multimodal features for body recognition by using appropriate pre-training deep learning neural network models and increasing the amount of data enhancement.
